PhD Student: Jamel Gafsi
Supervisor: Prof.Dr.Ernst W.Biersack
Key Words: Video Server Design, Striping, Reliability,
distributed Architecture.
Description:
Multimedia applications such as Video-On-Demand, Tele-Shoping, or distance learning require a video server for audio and video data. All these applications are very demanding in terms of storage capacity and I/O bandwidth. A video server must also meet the requirements that stem from the continuous nature of audio and video. It must guarantee the delivery of continuous media data in a timely fashion. Due to the large volume of the video data, a video server usually consists of a large number of disks and possibly multiple sever nodes.
We have designed a video server architecture called server array where each video is distributed over all server nodes, i.e. each node stores only a part of the video. The major advantage of the proposed architecture is its superior scalability. Independent of the demand distribution, the load will always be equally distributed over all video server nodes.
Our research focuses on modeling and designing a distributed and reliable video server. A challenging task thereby is to provide both, (i) performance (the number of concurrent clients and the start-up latency for new client request) by distributing the workload evenly among disks, and (ii) reliability by introducing redundant information to protect against component failures.
We have implemented a VOD prototype based on our research results. Both server and clients run on multiple platforms such as Sun machines running Solaris and Pc's running Windows NT. The server stores MPEG-I coded videos and serves a large number of clients, each of them consuming the MPEG streams at the rate of 25 frames/sec. The prototype is able to cope with many types of heterogeneity: the number of disks per node, their storage and I/O bandwidth capacity, the variety of the striping techniques used, the diversity of the reliability schemes supported (Error Correcting Codes (ECC) in software, or mirroring-based schemes).
We are now testing our video server on a high performance MAN (EuroSud 155) and studying its performance and quality of service.
Recently, we studied reliability on a video server from two different points of view:
We have compared different mirroring schemes in terms of their performance and reliability using discrete Markov models. Our results show that a mirroring scheme that copies data of a single disk onto a subset of all disks achieves the best trade-off between reliability and the number of concurrent streams supported.
We have compared mirroring- and parity-based schemes in
terms of performance and cost. Our results show that a server that uses
mirroring, although it doubles the storage requirement, has lower per stream
cost than a server that uses parity-based reliability.
Slides